test(SAMPLE-REASONING): pin the two qwen3 branches nothing caught, and correct a 2-core that was never possible (#605, #606)

[localai-bot]

Closes the three LOW findings accepted as follow-up when the recipe-sweep PRs
merged today. Issues: #605
(findings 1 and 2, the SAMPLE-REASONING W3 brick reviewed in #630) and
#606 (finding 3, the
serve-recipe-args spec reviewed in #624). No src/, include/ or behaviour
change — this is a test and a record.

1. #630 LOW-1 — a load-bearing branch with no test

src/vllm/parser/qwen3.cpp:55's reasoning-reopen guard (if (ps != npos && ps > p) return false;)
was correct but unpinned. Reproduced on unmodified main first: deleting it left
the focused suite fully green — 7 cases / 7 passed, 157 assertions / 0 failed,
Status: SUCCESS!, exit 0. All four existing tool-call assertions put <think>
BEFORE the <tool_call>, so ps > p never fires.

Added the one input that reaches it. With the assertion present and the guard
still deleted:

test_qwen3.cpp:273: ERROR: CHECK( p->is_reasoning_end("<tool_call>x<think>y") == false ) is NOT correct!
  values: CHECK( true == false )
[doctest] test cases:   8 |   7 passed | 1 failed | 0 skipped
[doctest] assertions: 164 | 163 passed | 1 failed |
[doctest] Status: FAILURE!

One failure, and it is the new assertion. (The doctest output above is verbatim
from the mutation run, so its line numbers are the ones the file had at that
moment; in the committed file the assertion is test_qwen3.cpp:288 and the
seed_oss case below runs from :303, failing at :317.) Head is the correct side: upstream's
backward token walk tests the reasoning-start id at every index before the
tool-call id (vllm/parser/qwen3.py:263-268), so it meets <think> first and
returns False.

2. #630 LOW-2 — the seed_oss thinking-off arm

Judged worth pinning rather than skipping. The seed_oss refactor is a fidelity
fix (class SeedOssParser(Qwen3Parser), seed_oss.py:24), and the reviewer
established it is invisible in production: OpenAIServingChat::MakeParserEngine
(serving_chat.cpp:580) takes the header default thinking=true, under which a
bare engine and Qwen3Parser are byte-identical. That is precisely the argument
for the test — the thinking-off arm is the only observer the invariant has, so
an unpinned inheritance that no production path exercises is exactly the kind a
later refactor drops silently. Driven through get_parser_engine("seed_oss") so
it pins the routing as well as the behaviour.

Mutation — route seed_oss back to the pre-#630 bare ParserEngine:

test_qwen3.cpp:316: ERROR: CHECK( content == Opt(output) ) is NOT correct!
[doctest] test cases:   8 |   7 passed | 1 failed | 0 skipped
[doctest] assertions: 164 | 163 passed | 1 failed |
[doctest] Status: FAILURE!

Coverage beyond upstream, recorded

test_qwen3_reasoning_parser.py has no qwen3 is_reasoning_end case at all
(the only one in the suite is the base family's,
test_base_thinking_reasoning_parser.py:111), and its THINKING_DISABLED_CASES
run with tool parsing suppressed. Both Qwen3Parser overrides therefore sit on
branches upstream's own fixtures cannot reach. The three affected TEST_CASEs are
written from scratch against the upstream source and disclosed in the test header
plus a new subsection of .agents/specs/reasoning-parsers.md § Tests to port,
which tabulates each case with the mutation that proves it.

Filed in the row spec rather than porting-inventory.md §9: §9 is the register
for implementation deviations forced by the no-Python constraint, and the
precedent for beyond-upstream test coverage is the row spec
(kv-persistence-lmcache.md:537, kv-events.md:196). Flagging the deviation
from the finding's wording explicitly rather than silently.

3. #624 LOW-1 — a wrong fact on main

.agents/specs/serve-recipe-args.md:278 and :284 said "2-core CI runner" for a
run a 2-core box cannot produce. Anchors re-verified in the tree, not trusted
from the finding (that is #632):

• tests/vt/test_cpu_threadpool.cpp:501 — if (cores < 4) { … return; }, so a
  box that produced a ratio has at least 4.
• :512-513 — fits = cores / 2, over = cores + 1; the recorded run reports
  2 and 5 threads, pinning cores == 4.

Both lines corrected, and the derivation added so the number now carries its own
proof. #631 is untouched — it was already retitled 4-core upstream of this;
this only makes the spec agree with it.

Gates

All observed in this worktree at 983b4aa4c + this diff:

gate	result
focused RED (guard mutated, assertion present)	exit 1, Status: FAILURE!, 1 failure, the new assertion
focused RED (seed_oss routing mutated)	exit 1, Status: FAILURE!, 1 failure, the new case
focused GREEN (tree restored)	exit 0, 8 cases / 8 passed, 164 assertions / 0 failed
full serial ctest --test-dir build	exit 0, 408/408 passed, 0 failed
scripts/agent-preflight.sh --staged	exit 0, all gates green
python3 scripts/check-windows-portability.py	exit 0
python3 scripts/check-agent-record.py	exit 0

Both mutated sources were restored byte-for-byte and re-verified by md5 and an
empty git diff -- src/. Every build's exit code was checked before its test run,
and each mutation was rebuilt rather than re-running a stale binary — the first
guard-deletion attempt in fact failed -Werror on an unused ps, which a
discarded build log would have hidden.

test_cpu_threadpool is green here (20-core box); its CI red is the pre-existing
#631, and windows-msvc-* red is the pre-existing #584. Neither is from this diff.

🤖 Generated with Claude Code